Deep-Feature-Consistent-VAE

Pytorch Implementation of Hou, Shen, Sun, Qiu, "Deep Feature Consistent Variational Autoencoder", 2016

Requirements

• Python 3.6.9
• matplotlib 3.2.1
• numpy 1.19.0
• pillow 7.2.0
• pytorch 1.5.1
• tensorboard 2.2.2
• torchvision 0.6.1

1) Installing Dependencies

pip3 install --upgrade pip
pip3 install -r requirements.txt

2) Train Vae

Either execute the (modified) example script, i.e.

cd examples
bash train_vae.sh

or

python3 train_vae.py --imgdir <path> --loss 0 -o pvae
python3 train_vae.py --imgdir <path> --loss 1 -o vae123
python3 train_vae.py --imgdir <path> --loss 2 -o vae345

3) Visualize Vae Reconstructions

Either execute the (modified) example script, i.e.

cd examples
bash visualize_vae.sh

or

python3 ../visualize_vae.py --vae pvae.pt vae123.pt vae345.pt --imgdir <path>

From top to bottom: original image, pvae, vae123, vae345

4) Visualize Linear Interpolation of Latent Space

Either execute the (modified) example script, i.e.

cd examples
bash visualize_latent_interpolation.sh

or

python3 ../visualize_latent_interpolation.py --vae vae123.pt --img_left <path1> --img_right <path2>

original α=0 ---------------------------------------------------------------------------------------------------> α=1 original

5) Visualize Facial Attribute Manipulation

Either execute the (modified) example script, i.e.

cd examples
bash visualize_facial_attribute_manipulation.sh

or

python3 ../visualize_facial_attribute_manipulation.py --vae vae123.pt --list_attr <path1> --attr <attr> --imgdir <path2> --img <path3>

Example adding eyeglasses: original α=0 ---------------------------------------------------------------------------------------------------> α=1